Causal inference from time series aims to determine which variables mechanistically drive which, but is undermined when measurements are collected far more slowly than the causal processes they record — as in functional MRI, where signals are sampled every one to two seconds while neural interactions unfold in milliseconds. This temporal undersampling distorts the recovered graph, collapsing indirect pathways into spurious direct edges and rendering many distinct causal structures equally consistent with the data. This dissertation develops a scalable, noise-tolerant framework for rate-agnostic causal structure learning, which recovers the causal graph at a system's native timescale without assuming the undersampling rate is known. We first recast the problem as constraint satisfaction in Answer Set Programming, yielding a solver that preserves prior theoretical guarantees while running about three orders of magnitude faster and scaling to graphs of over one hundred variables. We then show, counterintuitively, that deliberately measuring more slowly at coprime timescales can reduce structural uncertainty. Finally, we extend the framework to real neuroimaging through a noise-tolerant meta-solver that corrects undersampling artifacts and returns calibrated sets of directed graphs; applied to resting-state fMRI from the FBIRN cohort (151 controls, 160 individuals with schizophrenia), it recovers a directed connectome that reproduces canonical findings in oriented form and flags which edge directions are stable across timescales.